KR20200142374A - Method for selecting artificial intelligience model based on input data and disaply apparatus for performing the same method thereof - Google Patents

Abstract

An embodiment of the present invention discloses an AI model selection method based on input data to automatically select an AI model enabling accurate acquisition of a result corresponding to data to be tested, for example, a classification result of one of at least one class, among a plurality of AI models, and a method performing the same accordingly. The AI model selection method comprises the steps of: acquiring misclassified data, which is data corresponding to a misclassified result of training data sets; acquiring a learned additional AI model; and outputting misclassification probabilities for each of the plurality of AI models.

Description

AI model selection method based on input data and a display device that performs the method according to it {METHOD FOR SELECTING ARTIFICIAL INTELLIGIENCE MODEL BASED ON INPUT DATA AND DISAPLY APPARATUS FOR PERFORMING THE SAME METHOD THEREOF}

The disclosed embodiment relates to a method for selecting an AI model based on input data and a display device performing the method according to the method.

Specifically, the disclosed embodiment relates to a method for selecting an AI model based on input data for outputting a result corresponding to the input data, and a display device performing the method according to the method.

Artificial Intelligence (AI) system is a system in which a machine learns and judges itself, derives a desired result, or performs a target operation.

AI technology consists of machine learning, such as deep learning, and component technologies that utilize machine learning. AI technology is widely used in technical fields such as linguistic understanding, visual understanding, reasoning/prediction, knowledge expression, and motion control.

For example, AI technology can be used in the art of visual understanding and inference/prediction. Specifically, it is possible to implement a technology to analyze and classify input data using AI technology. That is, it is possible to implement a method and apparatus capable of obtaining a desired result by analyzing and/or classifying the input data.

Hereinafter, among various AI technologies, an AI technology that analyzes input data, recognizes an object included in the input data (eg, image), classifies the recognized object into at least one class, and outputs it. It will be described with an example.

These AI technologies can be implemented using algorithms. Here, an algorithm or a set of algorithms for implementing AI technology is called a neural network. Here, the neural network may receive input data, perform an operation for analysis and classification described above, and output result data. In this way, in order for the neural network to accurately output the result data corresponding to the input data, it is necessary to train the neural network. Here,'training' is a method of inputting various data with a neural network, analyzing the input data, classifying the input data, and/or extracting features necessary for generating result data from the input data. It can mean training a neural network so that it can discover or learn how to do it by itself. Here,'training' may be expressed as'learning' or'training' in Korean.

In addition, a set of algorithms for outputting output data corresponding to input data through the aforementioned neural network, software for executing the set of algorithms, and/or hardware for executing the set of algorithms are referred to as'AI models' (or'artificial). Intelligence model').

AI models can exist in many different forms. Specifically, there may be a plurality of AI models that receive an image and perform an operation of analyzing the input image to classify objects included in the image into at least one class.

Here, when each of the plurality of AI models generates output data corresponding to the input data, the accuracy of the output data may differ from each other in the plurality of AI models. In addition, output results from a plurality of AI models may vary according to input data. That is, when the same input data is input, the result data may have a true value in the first AI model, but the result data may have an incorrect value in the second AI model.

Among these various AI models, it is necessary to determine which model is used to input data and obtain results. Furthermore, there is a need to provide a method and apparatus for selecting an AI model that can increase processing accuracy of input data among AI models that exist in various forms.

An object of the disclosed embodiment is to provide an AI model selection method based on input data and a display device performing the method according to the input data, which enables one AI model to be automatically selected from among a plurality of AI models.

Specifically, the disclosed embodiment is a method of selecting an AI model based on input data and a display device performing the method according to the input data, which enables the selection of an AI model optimized for processing input data to be processed from among a plurality of AI models. It is for the purpose of providing.

In the method of selecting an AI model based on input data according to the disclosed embodiment, when each of a plurality of AI (Artificial Intelligence) models receives a training data set, classifies it into at least one class, and outputs a result, the training data Obtaining misclassified data, which is data corresponding to a result that is misclassified among the three; Learning the misclassified data from an additional AI model to obtain a learned additional AI model; Inputting first data into the learned additional AI model and outputting misclassification probabilities for each of the plurality of AI models; And selecting any one of the plurality of AI models based on the misclassification probabilities.

In addition, the selecting may include selecting a model corresponding to a misclassification probability having a smallest value among the misclassification probabilities among the plurality of AI models.

In addition, the learning may include learning from which of the plurality of AI models the misclassified data is classified data from the additional AI model receiving the misclassified data; And obtaining the learned additional AI model.

In addition, the learning may include the step of learning, by the additional AI model receiving the misclassification data, the characteristics of data and distribution of data that cannot be learned by each of the plurality of AI models. have.

In addition, the method of selecting an AI model based on input data according to the disclosed embodiment may further include receiving first data from the selected AI model and outputting a result of classifying it into the at least one category.

In addition, the additional AI model may be an AI model different from the plurality of AI models.

In addition, the misclassification probability may represent a probability that each of the plurality of AI models incorrectly classifies the input data.

In addition, the obtaining of the misclassification data may include receiving the training data set from each of the plurality of AI models, classifying the training data set into the at least one category, and outputting the result; And acquiring the misclassification data, which is data having an incorrect value, from among the training data set.

In addition, in the plurality of AI models, at least one of a hyper-parameter, a model architecture, a training technique, and a data set input for training may be formed differently from each other.

In addition, the AI model selection method based on the input data according to the disclosed embodiment is based on the misclassification probabilities, in each of the plurality of AI models in the ensemble AI model formed by combining the plurality of AI models. It may further include adjusting a plurality of weight values to be applied.

In addition, the adjusting of the plurality of weight values may include adjusting a plurality of weight values applied to each of the plurality of AI models so that the applied weight value decreases when the misclassification probability is high. I can.

In addition, the method of selecting an AI model based on input data according to the disclosed embodiment may further include inputting the first data into the ensemble AI model and outputting a final classification result corresponding to the first data.

In addition, the plurality of AI models may include an ensemble AI model formed by combining a plurality of AIs.

In addition, the plurality of AI models may include at least one on-device AI model and at least one server based AI model.

A display device according to the disclosed embodiment includes a display; And when executing at least one instruction, each of a plurality of AI (Artificial Intelligence) models receives a training data set, classifies it into at least one class, and outputs a result, the misclassified of the training data set. (misclassified) Acquires misclassified data that is data corresponding to the result, trains an additional AI model with the misclassified data to obtain a learned additional AI model, and inputs first data as the learned additional AI model And a processor configured to control to output misclassification probabilities for each of the plurality of AI models, and to select any one of the plurality of AI models based on the misclassification probabilities.

In the computer program product according to the disclosed embodiment, when each of a plurality of AI (Artificial Intelligence) models receives a training data set, classifies it into at least one class, and outputs a result, misclassification of the training data set Obtaining misclassified data, which is data corresponding to a misclassified result; Learning the misclassified data from an additional AI model to obtain a learned additional AI model; Inputting first data into the learned additional AI model and outputting misclassification probabilities for each of the plurality of AI models; And a recording medium storing a program for performing an operation of selecting one of the plurality of AI models based on the misclassification probabilities.

An AI model selection method based on input data according to the disclosed embodiment, a display device for performing the method according to the method, and a program product for performing the method according to the method for processing input data to be processed among a plurality of AI models. Optimized AI models can be automatically selected.

Accordingly, it is possible to increase the accuracy of the output result output as a result of the operation performed in the AI model.

1A is a diagram for describing an AI technology for generating output data corresponding to input data using a neural network.
1B is another diagram for explaining an AI technique for generating output data corresponding to input data using a neural network.
2A is a flowchart illustrating a method of selecting an AI model based on input data according to the disclosed embodiment.
2B is another flowchart illustrating a method of selecting an AI model based on input data according to the disclosed embodiment.
3 is a diagram illustrating a training data set used in the disclosed embodiment.
4 is a diagram for describing a classification operation performed in the disclosed embodiment.
5 is a diagram for explaining misclassification data obtained in the disclosed embodiment.
6 is a diagram for explaining a plurality of AI models used in the disclosed embodiment.
7 is a diagram for explaining a learning operation of an additional model performed in the disclosed embodiment.
8 is a diagram for explaining misclassification probabilities output from a learned additional model.
9 is a detailed flowchart illustrating a method of selecting an AI model based on input data according to the disclosed embodiment.
10 is another flowchart illustrating in detail a method of selecting an AI model based on input data according to the disclosed embodiment.
11 is another flowchart illustrating in detail a method of selecting an AI model based on input data according to the disclosed embodiment.
12 is a block diagram illustrating a display device according to the disclosed embodiment.
13 is a block diagram illustrating a display device according to the disclosed embodiment in detail.
14 is another block diagram specifically illustrating a display device according to the disclosed embodiment.
15 is another block diagram illustrating a display device according to the disclosed embodiment.
16 is a diagram illustrating a server and a display device for performing an AI model selection method based on input data according to the disclosed embodiment.
17 is a block diagram specifically showing the server illustrated in FIG. 16.
18 is a block diagram illustrating a display device communicating with a server.
FIG. 19 is a diagram for describing operations performed in the server and display device illustrated in FIG. 16.

Hereinafter, embodiments of the present invention will be described in detail with reference to the accompanying drawings so that those of ordinary skill in the art can easily implement the present invention. However, the present invention may be implemented in various different forms and is not limited to the embodiments described herein. In the drawings, parts irrelevant to the description are omitted in order to clearly describe the present invention, and similar reference numerals are assigned to similar parts throughout the specification.

Throughout the specification, when a part is said to be "connected" to another part, this includes not only "directly connected" but also "electrically connected" with another element interposed therebetween. . In addition, when a part "includes" a certain component, it means that other components may be further included rather than excluding other components unless specifically stated to the contrary.

Phrases such as "in some embodiments" or "in one embodiment" appearing in various places in this specification are not necessarily all referring to the same embodiment.

Some embodiments may be represented by functional block configurations and various processing steps. Some or all of these functional blocks may be implemented with various numbers of hardware and/or software components that perform specific functions. For example, the functional blocks of the present disclosure may be implemented by one or more processors or microprocessors, or may be implemented by circuit configurations for performing an intended function. In addition, for example, the functional blocks of the present disclosure may be implemented in various programming or scripting languages. The functional blocks may be implemented as an algorithm executed on one or more processors. In addition, the present disclosure may employ conventional techniques for electronic environment setting, signal processing, and/or data processing. Terms such as modules and configurations may be widely used, and are not limited to mechanical and physical configurations.

In addition, the connecting lines or connecting members between the components illustrated in the drawings are merely illustrative of functional connections and/or physical or circuit connections. In an actual device, connections between components may be represented by various functional connections, physical connections, or circuit connections that can be replaced or added.

In addition, the description of'at least one of A and B'means'A or B'or'A and B'.

In the disclosed embodiment, the AI model may be composed of a plurality of neural network layers. Each layer has a plurality of weight values, and a layer operation is performed through the operation result of a previous layer and a plurality of weights.

In addition, the neural network constituting the AI model may be a set of algorithms that learn a method of recognizing an object from an image input to the neural network based on artificial intelligence (AI) technology.

For example, a neural network can self-learn the types of data necessary to recognize an object from an image without (i) supervised learning with an image as an input value, and/or (ii) without a map. Based on unsupervised learning to discover a pattern or method for recognizing an object, it is possible to learn a method of recognizing an object from an image. As another example, the neural network may learn a method of recognizing an object from an image by using reinforcement learning using feedback on whether a result of recognizing an object according to learning is correct. As another example, the neural network may learn a method of recognizing an object from an image using semi-supervised learning.

In addition, a neural network performs an operation based on artificial intelligence (AI) technology, and may be referred to as an artificial neural network. In addition, the neural network may be a deep neural network (DNN) that performs an operation through a plurality of hidden layers.

Specifically, a neural network is classified as a deep neural network when the number of hidden layers, which is an inner layer that performs an operation, is plural, that is, when the depth of a neural network that performs an operation increases. I can. Examples of neural networks include CNN (Convolutional Neural Network), DNN (Deep Neural Network), RNN (Recurrent Neural Network), RBM (Restricted Boltzmann Machine), DBN (Deep Belief Network), BRDNN (Bidirectional Recurrent Deep Neural Network), and deep Q-networks (Deep Q-Networks), and the like, and the neural network in the present invention is not limited to the above-described example except for the case specified. In addition, the CNN neural network may be subdivided into a Deep Convolution Neural Network (DCNN) or a Capsnet neural network (not shown).

As described above, the'AI model' may refer to a neural network including at least one layer operating to receive input data and output a desired result. In addition, the'AI model' is an algorithm that outputs a desired result by performing an operation through a neural network, or a set of multiple algorithms, a processor for executing such an algorithm or a set thereof, and executing such an algorithm or a set thereof. It may mean software for, or hardware for executing such an algorithm or a set thereof.

In the disclosed embodiment, the AI model performs an operation through a neural network and outputs a result corresponding to the input data. The AI model used in the disclosed embodiment may be formed as a neural network for recognizing and classifying objects included in the input data by analyzing input data.

In addition, the AI model used in the disclosed embodiment may be formed as a neural network for analyzing input data and outputting a classification result classified according to a predetermined criterion.

The AI model used in the disclosed embodiment will be described below with reference to FIGS. 1A and 1B.

1A is a diagram for explaining an AI technology for generating output data corresponding to input data using a neural network. In FIG. 1A, a deep neural network (DNN) in which the hidden layer 120 is formed of three layers is illustrated.

Specifically, FIG. 1A shows an AI model that receives an image as an input, recognizes and classifies an object in a deep neural network, and outputs a result corresponding to the received image, and the AI model is formed by the deep neural network 101 The case is illustrated as an example.

Referring to FIG. 1A, the deep neural network 101 may be trained by receiving training data. In addition, the learned deep neural network 101 may perform an operation for object recognition. Here, the deep neural network 101 may be designed in a wide variety according to a model implementation method (eg, a convolution neural network (CNN), etc.), accuracy of results, reliability of results, and processing speed and capacity of a processor. .

The deep neural network 101 may perform an operation for object recognition, including the input layer 111, the hidden layer 120, and the output layer 130.

In addition, each of the plurality of layers forming the deep neural network 101 may include one or more nodes. For example, the input layer 111 may include one or more nodes (eg, 110) that receive data. In FIG. 1A, an example in which the input layer 111 includes a plurality of nodes is illustrated. In addition, each of the plurality of images obtained by scaling the input image 103 may be input to the plurality of nodes 110. Specifically, a plurality of images obtained by scaling the image 103 for each frequency band may be input to the plurality of nodes 110.

Here, two adjacent layers are connected by a plurality of edges (eg, 112) as shown. Since each node has a corresponding weight value, the deep neural network 101 can obtain output data based on a value obtained by calculating, for example, multiplying the input signal and the weight value. In FIG. 1A, a case in which nodes included in one layer are connected in a'full-connected' manner with nodes included in adjacent other layers as a whole is illustrated. However, nodes included in one layer may be partially connected to nodes included in other adjacent layers. In this case, at least one node included in one layer may not be connected to at least one node included in another adjacent layer.

The deep neural network 101 may be trained based on a plurality of training images, and may be constructed as an object recognition model for recognizing an object included in the image. Specifically, in order to increase the accuracy of the result output through the deep neural network 101, training is repeatedly performed from the output layer 130 to the input layer 111 based on a plurality of training images, and the output result The weight values can be modified to increase the accuracy of.

In addition, the deep neural network 101 having finally modified weight values may be used as an object recognition model. When the deep neural network 101 receives an image and is trained to recognize an object included in the image, when the deep neural network 101 receives the image 103, the deep neural network 101 analyzes the image 103 It is possible to output the result that the object contained within is'panda'.

1B is another diagram for explaining an AI technology for generating output data corresponding to input data using a neural network. In Fig. 1B, the same configuration as in Fig. 1A is shown using the same reference symbols.

Referring to FIG. 1B, the AI model used in the disclosed embodiment may be formed as a neural network for analyzing input data and outputting a classification result classified according to a predetermined criterion. Specifically, the AI model used in the disclosed embodiment may be formed of a deep neural network 101 trained to classify the type or type of input data by analyzing input data.

Input data input to the deep neural network 101 may be data of various types. Specifically, the input data may be image data as illustrated in FIG. 1A, and may be sound data, text data, or the like. In FIG. 1B, a case where the deep neural network 101 receives sound data, analyzes the input sound data, and outputs a classification result classified into at least one class will be described as an example.

Referring to FIG. 1B, the deep neural network 101 may receive sound data and analyze the input sound data through the hidden layer 120 included in the deep neural network 101. Further, the deep neural network 101 may output a result 190 of classifying the input sound data into at least one category or type, for example, speech, music, noise, etc. have.

For example, if the input sound data is voice data uttered by the user, the deep neural network 101 may output a classification result of'Speech'.

In addition, the result 190 may include a detailed classification result. For example, if the input sound data corresponds to music, it is possible to output the progress classified into sub-classes of music, for example, classical, jazz, pop, rap, and traditional music. will be.

1A and 1B, a result output from the deep neural network 101 may have an incorrect value. For example, the accuracy of the result output from the illustrated deep neural network 101 may be 90%. In this case, the probability that the output result is incorrect may be 10%.

AI models that receive inputs and output desired results can exist in various forms. Specifically, the AI model is in various forms depending on the architecture of the included neural network, the setting of applied hyper-parameters, the training data used to train the neural network, and the training technique. Can exist as

In the case of receiving the same image, at least one of the structure of the neural network, the setting of the hyper-parameter applied, the training data used to train the neural network, and the training technique are formed differently. Multiple AI models may have different result accuracy.

In the example shown in FIG. 1A, it is assumed that a first AI model, a second AI model, and a third AI model having different structures of neural networks that perform object recognition and classification operations of recognized objects exist. . In addition, it is assumed that the result accuracy of the first AI model is 60%, the result accuracy of the second AI model is 70%, and the result accuracy of the third AI model is 90%. Here, the result accuracy may be a value calculated by averaging the result accuracy of a plurality of inputs for which the corresponding AI model outputs the result.

In this case, when the same image (for example, 130) in which a panda is drawn in each of the first AI model, the second AI model, and the third AI model is input, the first AI model outputs a result of'panda' and , The second AI model may output a result of'dog', and the third AI model may output a result of'person'.

In the above-described example, the model having the highest result accuracy among the first AI model, the second AI model, and the third AI model is the third AI model. However, in the third AI model with the highest accuracy of the results,'person', which is an incorrect classification result, was output, and in the first AI model with the lowest accuracy of the result,'panda', which is the correct classification result, was output.

As in the example, even in the case of the model with the highest result accuracy, there may be a case in which an incorrect result is output because object recognition is not accurately performed for a certain image according to the characteristics of the input image.

As described above, the accuracy of the result may vary depending on which of the plurality of AI models acquires result data corresponding to the input data. Accordingly, there is a need to provide a method and an apparatus for allowing a user to automatically select an AI model with high accuracy for data to be tested for deriving a result from among a plurality of AI models.

Hereinafter, with reference to the accompanying drawings, embodiments for selecting an AI model capable of outputting an accurate result for input data that a user wants to test will be described.

The method according to the disclosed embodiment is an AI model selection method based on input data for selecting an AI model optimized for input data input for testing.

The method according to the disclosed embodiment (for example, the AI model selection method 200 based on the input data shown in FIG. 2A) includes at least one processor (or at least one processor) executing at least one instruction. It may be performed by an electronic device including a processor). Here, by executing at least one of the one or more instructions, the processor may execute operations included in the method of selecting an AI model based on input data according to the disclosed embodiment. In addition, the processor may be implemented in hardware or software. Specifically, it may include a graphic processor (Graphic Processing Unit, not shown) for processing a graphic corresponding to a video. The processor (not shown) may be implemented as a SoC (System On Chip) in which a core (not shown) and a GPU (not shown) are integrated. The processor (not shown) may include a single core, a dual core, a triple core, a quad core, and a multiple of the core.

Also, the above-described processor may be mounted on at least one of an electronic device (not shown) and a server (not shown). Accordingly, the method according to the disclosed embodiment (for example, the AI model selection method 200 based on the input data shown in FIG. 2A) is an electronic device (not shown) and a server (not shown) equipped with the above-described processor. It may be performed through at least one of.

Here, the electronic device may be an electronic device including a display. Specifically, electronic devices including a display include mobile phones, tablet PCs, digital cameras, camcorders, laptop computers, tablet PCs, desktops, e-book terminals, digital broadcasting terminals, PDAs (Personal Digital Assistants), and PMPs (Portables). Multimedia Player), navigation, MP3 player, wearable device, and the like. Further, the electronic device may be a fixed electronic device disposed at a fixed position or a mobile electronic device having a portable form, and may be a digital broadcast receiver capable of receiving digital broadcasts. Hereinafter, an electronic device including a display will be referred to as a'display device'.

In addition, the server may include a server, a server system, a server-based device, etc. that transmit and receive data to and from an electronic device through a communication network and process data. Hereinafter, for convenience of description, a server, a server system, a server-based device, etc. will all be referred to as a'server'.

Hereinafter, a method according to the disclosed embodiment will be described in detail with reference to FIGS. 2A to 11. In addition, the configuration of an apparatus capable of performing a method according to the disclosed embodiment will be described in detail with reference to FIGS. 12 to 19 below.

2A is a flowchart illustrating a method of selecting an AI model based on input data according to the disclosed embodiment. In the AI model selection method 200 based on input data according to the disclosed embodiment, each of a plurality of AI (Artificial Intelligence) models receives a training dataset and classifies the result into at least one class. When outputting, obtaining misclassified data, which is data corresponding to a misclassified result of the training data set (S210), by learning the misclassified data from an additional AI model, the learned additional AI model Acquiring (S220), inputting first data into the learned additional AI model, and outputting misclassification probabilities for each of the plurality of AI models (S230), and the misclassification probability And selecting one of the plurality of AI models (S240).

Hereinafter, an AI model selection method 200 based on input data according to the disclosed embodiment will be described in detail with reference to FIGS. 2A to 8.

Referring to FIG. 2A, in the AI model selection method 200 based on input data, when each of a plurality of AI (Artificial Intelligence) models receives a training data set, classifies it into at least one class, and outputs a result. , Misclassified data corresponding to a misclassified result of the training data set is obtained (S210).

Here, the plurality of AI models may be different AI models that perform a result derivation operation intended by the user. Alternatively, the plurality of AI models may be different AI models available for testing. Alternatively, the plurality of AI models may be a server or device that performs the AI model selection method 200 based on input data, or a plurality of AI models that can be provided or used in a system in which the server and the device are combined.

For example, the plurality of AI models may be different AI models that perform object recognition and classification operations. For example, if a user wants to analyze an object included in an input image and classify it as a dog or cat, a plurality of AI models classify it by performing object recognition and classification of the object (e.g., dog or cat). You can print the result.

In addition, the training data set refers to a set of input data to train each of the plurality of AI models. In order to generate the desired output by analyzing the input data by the AI model, it is necessary to find or learn a method for obtaining or extracting the desired output by receiving a plurality of data and analyzing the inputted data. You have to go through the training. In this case, the set of data input for training the AI model may be referred to as a'learning data set'. That is, the training data set may be a set of data input to the AI model when training an AI model.

The training data set and a plurality of AI models will be described in detail below with reference to FIGS. 3 to 6. In FIGS. 3 to 8, the same configuration is illustrated using the same reference numerals.

3 is a diagram illustrating a training data set used in the disclosed embodiment.

Referring to FIG. 3, the training data set may represent a set of a plurality of images input to each of a plurality of AI models.

In FIG. 3, a case where the training data set 300 is a set of images 301, 305, 311, 315 including any one of a dog and a cat as an object is illustrated as an example. Hereinafter, a case where the training data set 300 includes 100 images including either a dog or a cat object will be described as an example. In addition, the number of images included in the training data set 300 is only an example, and much more images may be used to train an actual AI model.

In addition, each of the plurality of AI models mentioned in S210 will be described as an example of an AI model that can operate as a classifier that recognizes and classifies objects included in an image using the training data set 300. . For example, after each of the plurality of AI models receives input data, they may output a classification result for at least one class corresponding to the input data. For example, in the case of an AI model that classifies and outputs according to objects included in input data, the AI model may receive input data and output a corresponding classification result.

4 is a diagram for describing a classification operation performed in the disclosed embodiment. In FIG. 4, an AI model that recognizes an object included in an input image and classifies and outputs the recognized object type as a dog or cat will be described as an example.

Referring to FIG. 4, each of a plurality of AI models may receive images included in the training data set 300 for training, and recognize and classify objects included in each of the input images. And, the classification result output by each of the plurality of AI models may be any one of at least one class.

In FIG. 4, the AI model 410 may be any one of a plurality of AI models.

In each of the plurality of AI models, for example, when the AI model 410 learns the training data set 300, images in which the same object is imaged may be distinguished and learned. Specifically, the AI model 410 sequentially learns the images (305, 315, etc.) of the dog imaged in the training data set 300, and then sequentially learns the images (301, 311, etc.) of the cat. You can learn.

The AI model 410 may output a classification result corresponding to the input data. 4, the AI model 410 receives input data (for example, an image 305), recognizes an object included in the image 305, and recognizes the first class (C1: class 1) or the first class. Any one of the 2 classes (C2: class 2) can be output. In FIG. 4, a case where the first class (C1: class 1) is a dog and the second class (C2: class 2) is a cat is illustrated as an example.

Assume that the AI model 410 receives an image 305 including a dog as an object. At this time, if the AI model 410 correctly recognizes and classifies the object included in the input image, the classification result of'dog' will be output, and the AI model 410 correctly recognizes the object included in the input image. And if the classification is not possible, a classification result of'cat' will be output.

Hereinafter, when the AI model 410 outputs an incorrect result, the input data will be referred to as'misclassified data'. For example, when the AI model 410 receives the image 305 and outputs the classification result as'cat', the image 305 may be included in'misclassified data'.

5 is a diagram for explaining misclassification data obtained in the disclosed embodiment.

In addition, in FIGS. 5 to 8, a case where the plurality of AI models mentioned in step S210 includes three AI models will be described as an example.

Referring to FIG. 5, a plurality of AI models trained using the training data set 300 may include a first AI model M1, a second AI model M2, and a third AI model M3. .

5, the first AI model M1 receives the training data set 300, learns each of a plurality of images included in the training data set 300, and performs object recognition and classification operations. I can. In the case of the first AI model (M1), 82% (511) of the classification result is correct, and 18% (512) of the case where the classification result is incorrect (misclassification). Accordingly, when the training data set 300 includes 100 images, the misclassification data collected by the first AI model M1 will be 18 images of the training data set 300.

The second AI model M2 may receive the training data set 300 and learn each of a plurality of images included in the training data set 300 to perform object recognition and classification operations. In the case of the second AI model (M2), 75% (521) of the classification result is correct, and 22% (522) when the classification result is incorrect (misclassification). Accordingly, when the training data set 300 includes 100 images, the misclassification data collected by the second AI model M2 will be 25 images of the training data set 300.

The third AI model M3 receives the training data set 300, learns each of a plurality of images included in the training data set 300, and performs object recognition and classification operations. In the case of the 3rd AI model (M3), 80% (531) of the classification result is correct, and 20% (532) of the incorrect classification result (misclassification). Here, the probability when the classification result is correct can be expressed as classification accuracy. Accordingly, when the training data set 300 includes 100 images, the misclassification data collected by the third AI model M3 will be 20 images of the training data set 300.

Referring to FIG. 5, the misclassification data collected in the first AI model M1 may be data corresponding to the misclassification probability (18%) of the training data set 300. In addition, the misclassification data collected by the second AI model M2 may be data corresponding to the misclassification probability (25%) of the training data set 300. In addition, the misclassification data collected by the third AI model M3 may be data corresponding to the misclassification probability (20%) of the training data set 300.

As illustrated in FIGS. 4 and 5, each of the plurality of AI models may have different classification accuracy and/or classification results.

Each of the plurality of AI models may have different classification accuracy even when the same training data set is input. In addition, each of the plurality of AI models may output different classification results even when the same training data set is input.

6 is a diagram for explaining a plurality of AI models used in the disclosed embodiment.

In the disclosed embodiment, the plurality of AI models include at least one of a neural network architecture, an applied hyper-parameter setting, training data used to train a neural network, and a training technique. It can be formed differently.

Referring to FIG. 6, a plurality of AI models, as expressed in block 600, include the structure of a neural network, setting of applied hyper-parameters, and training data used to train the neural network. , At least one of the training techniques may be formed differently.

For example, the first AI model M1 may be a model 611 trained by a first training data set (Dataset 1), and the second AI model (M2) is a second training data set (Dataset 2). ), and the third AI model M3 may be a model 613 trained by the third training data set (Dataset 3). Here, the first training data set (Dataset 1), the second training data set (Dataset 2), and the third training data set (Dataset 3) represent different sets of images.

As another example, the first AI model M1 may be a model 621 having a first hyper parameter (eg, hyper-parameter vector 1), and the second AI model M2 is a second hyper parameter (Eg, hyper-parameter vector 2) may be a model 622 having, and the third AI model (M3) is a model 623 having a third hyper parameter (eg, hyper-parameter vector 3). ) Can be. In addition, a first hyper parameter (eg, hyper-parameter vector 1), a second hyper parameter (eg, hyper-parameter vector 2), and a third hyper parameter (eg, hyper-parameter vector 3) are They are set to different values.

Here, the hyper parameter may include parameters such as the number of hidden layers in the neural network forming the AI model, a learning rate, a batch size, a loss function, and a weight decay rate.

As another example, the first AI model M1 may be a model 631 having a first structure (Deep-architecture 1), and the second AI model M2 is a second structure (Deep-architecture 2). It may be a model 632 having, and the third AI model M3 may be a model 633 having a third structure (Deep-architecture 3). Here, the structure represents the structure of the AI model, and the first structure (Deep-architecture 1), the second structure (Deep-architecture 2), and the third structure (Deep-architecture 3) are different from each other.

As another example, the first AI model M1 may be a model 641 formed by a first learning technique (eg, domain adaptation technique), and the second AI model M2 is a second learning technique ( For example, it can be a model 642 having a multi-tasking learning technique), and the third AI model M3 will be a model 633 having a third learning technique (eg, methodology XYZ technique). I can. That is, the first AI model M1, the second AI model M2, and the third AI model M3 may be formed by different training techniques.

As described above, at least one of the structure of the neural network, the setting of the hyper-parameter to be applied, the training data used to train the neural network, and the training technique are formed differently. When there are two AI models, even if the same input data is input to each of the plurality of AI models, the recognition accuracy of the plurality of AI models may be different from each other. In addition, since AI models can be designed in a wide variety, there may be various AI models that perform the same operation.

For example, for specific input data, the first AI model M1 is accurately recognized and classified, but the second AI model M2 is not recognized.

In the disclosed embodiment, accordingly, a method and an apparatus for selecting an AI model optimized for input data to be tested by a user from among various different AI models are disclosed.

Referring back to FIG. 2A, the method 200 for selecting an AI model based on input data learns the misclassified data from an additional AI model, following step S210, to obtain a learned additional AI model (S220). Specifically, in step S220, in the additional AI model receiving the misclassified data, learning whether the misclassified data is data classified from among a plurality of AI models (step not shown) and the result of the training, the learned It may include obtaining an additional AI model (step not shown).

Here, the additional AI model is an AI model different from the plurality of AI models mentioned in step S210. In addition, the additional AI model may mean an AI model to be trained by misclassification data. The training operation of the additional AI model will be described in detail below with reference to FIG. 7.

7 is a diagram for explaining a learning operation of an additional model performed in the disclosed embodiment.

Referring to FIG. 7, a misclassified dataset 710 for training an additional AI model M0 refers to a set of misclassified data, which is data misclassified in each of a plurality of AI models. .

As described with reference to FIG. 5, for example, when the training data set 300 includes 100 images. Then, the misclassified dataset 710 is from 18 images collected from the first AI model (M1), 25 images collected from the second AI model (M2), and from the third AI model (M3). It can be a collection of 20 images collected.

For example, the additional AI model M0 may sequentially train misclassification data collected from each of a plurality of models.

Specifically, the additional AI model M0 trains 18 images misclassified in the first AI model M1, and subsequently trains 25 images misclassified in the second AI model M2 ( training), and subsequently, it is possible to train 20 images misclassified in the third AI model (M3).

Specifically, the additional AI model M0 learns the misclassified data 710 to determine whether the input data is data classified from any of a plurality of AI models (eg, M1, M2, and M3). You can learn it yourself. That is, the additional AI model (M0) uses the misclassification data 710 to train whether the input data is data classified from among a plurality of AI models (eg, M1, M2, and M3). can do.

Specifically, the additional AI model M0 receives the misclassification data 710 and learns it, so that the characteristics of data that cannot be learned by each of a plurality of AI models (eg, M1, M2, and M3) and You can learn the distribution of data.

For example, the additional AI model M0 trains 18 images misclassified in the first AI model M1, so that the first AI model M1 recognizes and recognizes images having certain image characteristics. It is possible to learn by itself whether it is impossible to classify or what image characteristics of an image that the first AI model M1 cannot recognize and/or classify has. Here, the image characteristics may refer to features expressing an object such as a resolution of an image, a method of expressing an object, a pixel value, a characteristic of an edge included in an image, and a color of an image.

In addition, the additional AI model M0 trains 25 images misclassified in the second AI model M2, so that the second AI model M2 recognizes and/or recognizes an image having certain image characteristics. Whether the second AI model M2 recognizes and/or does not classify, it is possible to learn by itself what image characteristics have. Subsequently, the additional AI model M0 is trained on 20 images misclassified in the third AI model M3, so that the third AI model M3 recognizes and/or recognizes images having certain image characteristics. Alternatively, it is possible to learn by itself whether it is impossible to classify or what image characteristics of an image that the third AI model M3 cannot recognize and/or classify has.

Referring back to FIG. 2A, in the AI model selection method 200 based on input data, if the first data is input to the additional AI model M0 learned in step S220 following step S210, each of the plurality of AI models Misclassification probabilities of is output (S230).

Here, the first data is data input to the learned additional AI model MO. Specifically, the first data may be a target for outputting a result through the learned additional AI model MO.

For example, when a user wants to test an object included in a certain image using an AI model, an image to be tested may be the first data. That is, the first data may be test data.

The operation of step S230 will be described in detail with reference to FIG. 8 below.

8 is a diagram for explaining misclassification probabilities output from a learned additional model.

Referring to FIG. 8, the learned additional AI model (M0) 801 receives the first data 801, and the'misclassification probability', which is the probability that the first data will be misclassified in each of the plurality of AI models. Probability)' can be displayed. Here, the total sum of the misclassification probability corresponding to each of the plurality of AI models may be 1 or 100%.

As described above, in step S220, the additional AI model M0 receives the misclassification data 710 collected from a plurality of AI models (eg, M1, M2, M2) and learns it. Accordingly, the learned additional AI model M0 knows the characteristics of data that cannot be learned by each of the plurality of AI models (eg, M1, M2, and M3) and the distribution of the data. Therefore, when the learned additional AI model M0 receives the first data, it analyzes the data characteristics and data distribution of the first data 801, and the first data 801 having the analyzed data characteristics and data distribution is It is possible to calculate the probability of misclassification in each of a plurality of AI models (eg, M1, M2, and M3).

8, when a plurality of AI models include a first AI model (M1), a second AI model (M2), and a third AI model (M3), the first data 801 is the first AI model. The probability of misclassification in the model M1 is 0.7 (or 70%) 810, and the probability that the first data 801 is misclassified in the second AI model M2 is 0.2 (or 20%) 820 ), and, a probability that the first data 801 is misclassified in the third AI model M3 may be 0.1 (or 10%) 830.

That is, the class classified by the learned additional AI model M0 will be a misclassification probability value corresponding to each of a plurality of AI models (e.g., M1, M2 and M3) and a plurality of AI models. I can. For example, in the class classified by the trained additional AI model (M0), the first class (C1: class 1) 811 is the misclassification probability 810 in the first AI model (M1). The second class (C2: class 2) (821) becomes the misclassification probability 820 in the second AI model (M2), and the third class (C3: class 3) 831 is the third AI It may be a misclassification probability 830 in the model M3.

Referring again to FIG. 2A, the AI model selection method 200 based on input data selects any one of the plurality of AI models mentioned in step S210 based on the misclassification probabilities obtained in step S230 (S240). ).

Specifically, the AI model selection method 200 based on the input data selects an AI model (any one of a plurality of AI models) corresponding to the misclassification probability having the smallest value among the misclassification probabilities obtained in step S230. I can.

Referring to FIG. 8, the misclassification probabilities output by the learned additional AI model M0 are 0.7 (810), 0.2 (820), and 0.1 (830), the smallest of which is 0.1 (830). . Accordingly, the AI model selection method 200 based on the input data uses the third AI model M3, which is an AI model corresponding to the smallest probability value 0.1 830 among the plurality of AI models M1, M2, M3. You can choose.

Specifically, the AI model corresponding to the probability of having the smallest value among the misclassification probabilities output by the learned additional AI model (M0) means the AI model with the lowest probability of misclassifying the first data 801. do.

In addition, the AI model most optimized for the first data 801 may be said to be an AI model with the lowest probability of misclassifying the first data 801. Therefore, in the disclosed embodiment, by selecting any one of the plurality of AI models based on the probability of misclassification from the plurality of AI models, among the plurality of learned AI models, data to be tested (specifically, the first 1 The AI model that is least likely to misclassify the data (801)) can be automatically selected.

2B is another flowchart illustrating a method of selecting an AI model based on input data according to the disclosed embodiment. In FIG. 2B, the same and corresponding configurations as in FIG. 2A are illustrated using the same reference symbols. Therefore, in describing the method 200 for selecting an AI model based on the input data shown in FIG. 2B, the overlapping description as in FIG. 2A will be omitted.

Referring to FIG. 2B, the AI model selection method 200 may further include a step S250 in addition to the steps shown in FIG. 2A.

Referring to FIG. 2B, the AI model selection method 200 according to the disclosed embodiment may repeatedly perform operations S230 and S240 whenever first data is updated.

Specifically, the AI model selection method 200 may determine whether new or additional first data exists (S250), and may repeatedly perform operations S230 and S240 according to the determination result of step S250. In addition, the'new first data' described in step S250 of FIG. 2B is data different from the first data previously input for testing, and may mean all data input to an additional AI model for testing.

That is, the AI model selection method 200 repeatedly performs operations S230 and S240 based on the new first data so that when the first data is newly input, an AI model optimized for the newly input first data can be newly selected. Can be done.

In addition, the AI model selection method 200 according to the disclosed embodiment may repeatedly perform operation S230 and operation S240 on each of the plurality of data to be tested when there are a plurality of first data to be tested. have. For example, data that the user wants to test may exist as a'set', which is a plurality of data sets instead of one data. In this case, all data included in the test data set may be used as'first data'.

For example, when first to tenth test data exists in the test data set, the AI model selection method 200 may perform operations S230 and S240 by inputting the first test data as an additional AI model. . Then, a predetermined AI model corresponding to the first test data may be selected. Subsequently, the AI model selection method 200 may perform operations S230 and S240 by inputting second test data as an additional AI model. Then, a predetermined AI model corresponding to the second test data may be selected. In this manner, operations S230 and S240 may be repeatedly performed for each of the third to tenth test data that follow. Accordingly, for each of the first to tenth test data, an optimized AI model may be selected.

As described above, in the disclosed embodiment, a result corresponding to the data to be tested (specifically, the first data 801) among a plurality of AI models that exist in a wide variety of forms without manual manipulation by the user ( For example, it is possible to automatically select or detect an AI model that enables the most accurate acquisition of the classification result of at least one of the classes). Accordingly, it is possible to increase the computational accuracy (result accuracy) of the AI model for the data to be tested.

9 is a detailed flowchart illustrating a method of selecting an AI model based on input data according to the disclosed embodiment.

In the disclosed embodiment, the plurality of AI models trained by receiving the training data set described in step S210 include at least one on-device AI model and at least one server based AI model. I can.

Here, the on-device AI model refers to an AI model mounted in the electronic device when the AI model is mounted in an electronic device (not shown) other than a server. In addition, the server-based AI model refers to an AI model installed in the server when the AI model is mounted in the server (not shown).

Considering the size and performance of memory and processors, in general, in the case of the on-device AI model, memory usage and data processing capacity are formed smaller than that of the server-based AI model. Therefore, compared to the case of using the server-based on-device AI model, using the on-device AI model will allow data to be processed at a faster rate.

Also, depending on the input data to be tested, the output accuracy of the on-device AI model may have a higher value than the output accuracy of the server-based AI model. Therefore, in the disclosed embodiment, based on the misclassification probabilities described in step S230, one of an on-device AI model and a server-based AI model may be selected and used.

Referring to FIG. 9, steps S905 and S920 may correspond to steps S230 and S240, respectively.

Specifically, the'additional AI model' shown in step S910 is collected from a plurality of AI models including at least one on-device AI model and at least one server based AI model. This is an additional AI model trained using misclassified data, and may correspond to the additional AI model acquired in step S220.

Referring to FIG. 9, first data is input as the learned additional AI model M0, and a plurality of AI models (specifically, at least one on-device AI model and at least one server-based ( Server based) AI model) misclassification probabilities for each are output (S915). In the example shown in FIG. 9, the plurality of AI models include a first on-device AI model (On-Device Model 1), a second on-device AI model (On-Device Model 2), and a first server-based AI model ( Server1 (Model 1)), and a second server-based AI model (Server2 (Model 2)).

If the misclassification probability for the on-device AI model is smaller than the misclassification probability for the server-based AI model based on the misclassification probabilities obtained in step S915, the process proceeds to step S930.

Subsequently, considering the misclassification probabilities for each of the on-device AI models (S935), if the misclassification probability for the first on-device AI model 1 has the smallest value, the first on-device The AI model (On-Device Model 1) is selected (S937). In addition, if the probability of misclassification for the second on-device AI model 2 has the smallest value, the second on-device AI model 1 is selected (S939).

If the misclassification probability for the server-based AI model is smaller than the misclassification probability for the on-device AI model based on the misclassification probabilities obtained in step S915, the process proceeds to step S950.

Subsequently, considering the misclassification probabilities for each of the server-based AI models (S955), if the misclassification probability for the first server-based AI model (Server1 (Model 1)) has the smallest value, the first server-based The AI model (Server1 (Model 1)) is selected (S957). And, if the misclassification probability for the second server-based AI model (Server 2 (Model 2)) has the smallest value, the second server-based AI model (Server 2 (Model 2)) is selected (S959).

As described above, depending on the input data to be tested, the output accuracy of the on-device AI model may have a higher value than the output accuracy of the server-based AI model. Therefore, in the disclosed embodiment, when the on-device AI model has a lower probability of misclassification compared to the server-based AI model, the on-device AI model is selected and used to obtain output data corresponding to the input data. have. Then, it is possible to obtain output data by performing AI calculations quickly and accurately using an on-device AI model that operates using relatively little memory.

10 is another flowchart illustrating in detail a method of selecting an AI model based on input data according to the disclosed embodiment. In FIG. 10, the same and corresponding configurations as in FIG. 2A are illustrated using the same reference symbols. Therefore, in describing the method 1000 for selecting an AI model based on input data shown in FIG. 10, descriptions overlapping with those in FIGS. 1 to 9 will be omitted.

Referring to FIG. 10, the method 1000 for selecting an AI model based on input data may further include a step S1005 of learning each of the plurality of AI models using a training data set before step S210.

Then, as a result of performing step S1005, misclassification data is obtained from each of the plurality of AI models (S210).

In addition, the AI model selection method 1000 based on the input data may further include, after step S240, outputting a result corresponding to the first data by using the selected AI model (S1050). That is, the first data may be input from the selected AI model and a corresponding classification result may be output. Here, at least a class included in the classification result output in step S1050 includes the same class as the classification result output from each of the plurality of AI models mentioned in step S210. Specifically, each of the plurality of AI models mentioned in step S210 recognizes and classifies the object of the input image, and outputs a result indicating whether it is the first class (C1: class 1) or the second class (C2: class 2). Then, in step S1050, a result indicating whether the first class (C1: class 1) or the second class (C2: class 2) may be output may be output.

Specifically, when the first data is input to the selected AI model, the selected AI model may analyze the first data and perform a neural network operation to output a corresponding classification result. For example, if the selected AI model is an AI model that performs a classification operation shown in FIG. 4, the selected AI model recognizes and classifies the object included in the input first data, and the classification result (e.g., individual Or cat).

As described above, the disclosed embodiment increases the accuracy of the output result by selecting and using the AI model with the lowest probability of misclassifying the data to be tested (the'first data'), and generating the output result. I can make it.

11 is another flowchart showing in detail a method of selecting an AI model based on input data according to the disclosed embodiment. In FIG. 11, the same and corresponding configurations as in FIG. 2A are illustrated using the same reference symbols. Therefore, in describing the AI model selection method 1100 based on the input data shown in FIG. 11, descriptions overlapping with those in FIGS. 1 to 9 will be omitted.

Following step S240, the AI model selection method 1100 based on the input data includes each of the plurality of AI models in an ensemble AI model formed by combining the plurality of AI models based on misclassification probabilities. It may further include adjusting a plurality of weight values applied to (S1150).

Here, the ensemble AI model means an AI model formed by combining a plurality of AI models. A plurality of AI models may constitute an ensemble AI model through serial or parallel, or a combination of serial and parallel forms.

Specifically, in step S1150, a plurality of weight values applied to each of the plurality of AI models may be adjusted so that the applied weight value decreases when the misclassification probability is high. Specifically, among the plurality of AI models, the AI model with the highest misclassification probability applies the smallest weight value, so that the AI model with the highest misclassification probability in the ensemble AI model contributes to the formation of output data. can do. Accordingly, the output accuracy of the ensemble AI model can be increased.

For example, in step S1150, a plurality of AI models applied to each of the plurality of AI models are inversely proportional to a misclassification probability corresponding to each of the plurality of AI models and a weight value applied to each of the plurality of AI models. The weight values of the dogs can be adjusted.

As another example, step S1150 extracts a misclassification probability that is equal to or greater than a certain threshold from among misclassification probabilities corresponding to each of a plurality of AI models, and calculates the AI model corresponding to the extracted misclassification probability. You could set the weight to 0. Here, the AI model to which the weight value of 0 is applied cannot contribute to the formation of output data in the ensemble AI model.

That is, if the misclassification probability exceeds a certain threshold, it can be determined that the AI model does not accurately perform the AI operation to generate the result data. Therefore, in this case, the weight value may be set to 0 so that the AI model does not contribute to the formation of the output data.

When the adjusted weight values are applied to a plurality of AI models, and the ensemble AI model is composed of the applied AI models, the probability of misclassification of the input data can be lowered and the accuracy of the result can be improved accordingly.

Following step S1150, the AI model selection method 1100 based on input data inputs the first data into the ensemble AI model having an adjusted weight value, and outputs a final classification result corresponding to the first data. It may further include step S1160.

The AI model selection method (specifically, 200, 1000, or 1100) based on input data according to the disclosed embodiment described with reference to FIGS. 1A to 11 is at least one of a display device and a server to be described with reference to FIGS. 12 to 19. It can be done through one. In Figs. 12 to 19, the same configuration is illustrated using the same reference numerals.

Specifically, the AI model selection method (specifically, 200, 1000, or 1100) based on input data according to the disclosed embodiment described with reference to FIGS. 1A to 11 may be performed through a display device.

Display devices include mobile phones, tablet PCs, digital cameras, camcorders, laptop computers, tablet PCs, desktops, e-book terminals, digital broadcasting terminals, PDAs (Personal Digital Assistants), PMPs (Portable Multimedia Players), navigation systems, and MP3 devices. It can be implemented in various forms such as a player, a wearable device, and the like. In addition, the display device may be a fixed display device disposed at a fixed position or a portable display device having a portable form, and may be a digital broadcast receiver capable of receiving digital broadcasts.

12 is a block diagram illustrating a display device according to the disclosed embodiment.

Referring to FIG. 12, the display device 1200 includes a processor 1210 and a display 1220.

The processor 1210 executes at least one instruction and controls the following operations to be performed. Specifically, when the processor 12100 executes at least one instruction, each of a plurality of AI (Artificial Intelligence) models receives a training data set, classifies it into at least one class, and outputs a result, the learning Misclassified data, which is data corresponding to the misclassified result of the data set, is acquired, and an additional AI model is trained with the misclassified data to obtain a learned additional AI model, and the learned additional AI model is used. Control to output misclassification probabilities for each of the plurality of AI models by inputting first data, and to select any one of the plurality of AI models based on the misclassification probabilities. It works.

The processor 1210 may be configured with one or a plurality of processors. In this case, one or more processors may be a general-purpose processor such as a CPU or an AP, a graphics dedicated processor such as a GPU, or an artificial intelligence dedicated processor such as an NPU.

Specifically, the processor 1210 may control to process input data according to a predefined operation rule or an artificial intelligence model stored in a memory (not shown). A predefined motion rule or AI model is characterized by being created through training.

Here, being made through learning or being trained means that an artificial intelligence model with a desired characteristic is created by applying a learning algorithm to a plurality of training data. Such learning may be performed in the device itself performing artificial intelligence according to the present invention, or may be performed through a separate server/system.

Here, the learning algorithm is a method in which a predetermined target device (eg, a robot) is trained using a plurality of pieces of learning data so that a predetermined target device can make a decision or make a prediction by itself. Examples of the learning algorithm include supervised learning, unsupervised learning, semi-supervised learning, or reinforcement learning, and the learning algorithm in the present invention is specified. It is not limited to the above-described example except for.

Specifically, the processor 1210 may generate output data corresponding to input data through an AI model that performs an operation through a neural network described with reference to FIG. 1A. Specifically, the processor 1210 may perform an operation through at least one of the aforementioned plurality of AI models, additional AI models, and ensemble AI models.

Also, the processor 1210 may control operations performed by the display apparatus 1200 and may control other components included in the display apparatus 1200 to perform a predetermined operation.

The display 1220 outputs an image. Specifically, the display 1220 may output an image corresponding to the video data through an internally included display panel (not shown) so that the user can visually recognize the video data. In the disclosed embodiment, the display 1220 may output a user interface screen including output data generated by the processor 1210 through the AI model.

13 is a block diagram illustrating a display device according to the disclosed embodiment in detail.

Referring to FIG. 13, the display device 1300 may further include a communication unit 1330 compared to the display device 1200.

The communication unit 1330 may transmit and receive data through a communication network with at least one external device under the control of the processor 1210. Here, the external device may be a server (not shown), a mobile device (not shown), or a physically classified electronic device (not shown).

Specifically, the communication unit 1330 may include at least one communication module (not shown) for performing wired or wireless communication with an external device according to a predetermined wired or wireless communication standard. Specifically, the communication unit 1330 may include a communication module that performs wireless communication according to a communication standard such as a Bluetooth communication module, a wireless LAN, and wired Ethernet. Here, the wireless LAN communication module (not shown) may include a Wi-Fi communication module (not shown) for performing wireless communication according to the Wi-Fi communication standard.

Hereinafter, a case in which an external device that communicates with the display device 1300 is a server will be described as an example.

14 is another block diagram specifically illustrating a display device according to the disclosed embodiment.

Referring to FIG. 14, compared to the display device 1300, the display device 1400 may further include at least one of a memory 1440 and a user interface 1450.

The memory 1440 may store at least one instruction. Also, the memory 1440 may store at least one program executed by the processor 1210. Also, the memory 1440 may store at least one AI model described above. Specifically, the memory 1440 may store at least one of the aforementioned plurality of AI models, additional AI models, and ensemble AI models.

In addition, the memory 1440 is a flash memory type, a hard disk type, a multimedia card micro type, and a card type memory (eg, SD or XD memory). , RAM (Random Access Memory) SRAM (Static Random Access Memory), ROM (ROM, Read-Only Memory), EEPROM (Electrically Erasable Programmable Read-Only Memory), PROM (Programmable Read-Only Memory), magnetic memory, It may include at least one type of storage medium among magnetic disks and optical disks.

The user interface 1450 may receive a user input for controlling the display device 1400. The user interface 1450 includes a touch panel that senses a user's touch, a button that receives a user's push operation, a wheel that receives a user's rotation operation, a keyboard (key board), and a dome switch. It may include a user input device, but is not limited thereto.

In addition, the user interface 1450 may include a voice recognition device (not shown) for voice recognition. For example, the voice recognition device may be a microphone, and the voice recognition device may receive a user's voice command or voice request. Accordingly, the processor 1210 may control an operation corresponding to a voice command or voice request to be performed.

Further, the user interface 1450 may include a motion detection sensor (not shown). For example, a motion detection sensor (not shown) may detect movement of the display apparatus 1300 and receive the detected movement as a user input. In addition, the above-described speech recognition device (not shown) and motion detection sensor (not shown) may be included in the display device 1400 as modules independent from the user interface 1450, not in a form included in the user interface 1450. There will be.

15 is another block diagram illustrating a display device according to the disclosed embodiment.

The display device 1500 may correspond to the display device 1200, 1300, or 1400 according to the exemplary embodiment described with reference to FIGS. 12 to 14. In addition, the display apparatus 1500 may perform an AI model selection method based on input data according to the disclosed embodiment described with reference to FIGS. 1A through 11 through the display apparatus 1500 shown in FIG. 15. Therefore, in describing the display apparatus 1500, descriptions overlapping with those in FIGS. 1A to 14 will be omitted.

Referring to FIG. 15, the display apparatus 1500 includes a video processing unit 1510, a display unit 1515, an audio processing unit 1520, an audio output unit 1525, a power supply unit 1530, a tuner unit 1540, and a communication unit ( 1550), a sensing unit (not shown), an input/output unit 1570, a control unit 1580, and a memory 1590.

Here, the controller 1580 may correspond to the processor 1210 shown in FIG. 12. The communication unit 1550, the display unit 1515, and the memory 1590 of the display apparatus 1500 may correspond to the communication unit 1330, the display 1220, and the memory 1440 illustrated in FIG. 14, respectively. Accordingly, in describing the display apparatus 1500 illustrated in FIG. 15, descriptions overlapping with those of the display apparatus 1200, 1300, or 1400 described above will be omitted.

The video processing unit 1510 processes video data received by the display apparatus 1500. The video processing unit 1510 may perform various image processing such as decoding, scaling, noise filtering, frame rate conversion, and resolution conversion on video data.

The controller 1580 receives a request for recording video data processed by the video processing unit 1510, encrypts the video data, and performs a memory device (not shown) included in the controller 1580 or the memory 1590, for example. , It will be possible to control to be written to RAM (not shown).

In the disclosed embodiment, at least one processor (not shown) included in the controller 1580 executes at least one instruction, so that each of a plurality of AI (Artificial Intelligence) models receives a training data set and receives at least one class. Obtaining misclassified data, which is data corresponding to a misclassified result among the training data set, when the result is output by classifying it as (class) (S210), learning the misclassified data from an additional AI model Thus, obtaining the learned additional AI model (S220), inputting first data into the learned additional AI model, and outputting misclassification probabilities for each of the plurality of AI models ( S230) and, based on the misclassification probabilities, the execution of operations corresponding to the step S240 of selecting any one of the plurality of AI models may be controlled.

The display unit 1515 displays a video included in the broadcast signal received through the tuner unit 1540 on the screen under the control of the controller 1580. In addition, the display unit 1515 may display content (eg, video) input through the communication unit 1550 or the input/output unit 1570.

Also, the display unit 1515 may output an image stored in the memory 1590 under the control of the controller 1580. In addition, the display unit 1515 is a voice UI (User Interface: including, for example, a voice command guide) for performing a voice recognition task corresponding to voice recognition or a motion recognition task corresponding to motion recognition. A motion UI (eg, including a user motion guide for motion recognition) may be displayed.

The audio processing unit 1520 processes audio data. The audio processing unit 1520 may perform various processing such as decoding, amplification, noise filtering, and the like for audio data. Meanwhile, the audio processing unit 1520 may include a plurality of audio processing modules to process audio corresponding to a plurality of contents.

The audio output unit 1525 outputs audio included in the broadcast signal received through the tuner unit 1540 under the control of the controller 1580. The audio output unit 1525 may output audio (eg, voice, sound) input through the communication unit 1550 or the input/output unit 1570. Also, the audio output unit 1525 may output audio stored in the memory 1590 under the control of the controller 1580. The audio output unit 1525 may include at least one of a speaker 1526, a headphone output terminal 1527, or an S/PDIF (Sony/Philips Digital Interface: output terminal 1528). The audio output unit 1525 may include A combination of a speaker 1526, a headphone output terminal 1527, and an S/PDIF output terminal 1528 may be included.

The power supply unit 1530 supplies power input from an external power source to the components 1510 to 1590 inside the display apparatus 1500 under the control of the controller 1580. In addition, the power supply unit 1530 may supply power output from one or more batteries (not shown) located inside the display apparatus 1500 to the internal components 1510 to 1590 under the control of the controller 1580. I can.

The tuner unit 1540 tunes only the frequency of a channel desired to be received by the display apparatus 1500 among many radio components through amplification, mixing, resonance, etc. of a broadcast signal received by wire or wirelessly. It can be selected by (tuning). The broadcast signal includes audio, video, and additional information (eg, Electronic Program Guide (EPG)).

The tuner unit 1540 includes a user input (for example, a control signal received from an external control device (not shown), for example, a remote controller, for example, a channel number input, a channel up-down (up-down)). down) A broadcast signal may be received in a frequency band corresponding to a channel number (eg, cable broadcast No. 506) according to input and channel input from the EPG screen.

The tuner unit 1540 may receive broadcast signals from various sources such as terrestrial broadcasting, cable broadcasting, satellite broadcasting, and Internet broadcasting. The tuner unit 1540 may also receive a broadcast signal from a source such as analog broadcast or digital broadcast. The broadcast signal received through the tuner unit 1540 is decoded (eg, audio decoding, video decoding, or additional information decoding) to be separated into audio, video, and/or additional information. The separated audio, video, and/or additional information may be stored in the memory 1590 under the control of the controller 1580.

The tuner unit 1540 of the display apparatus 1500 may be one or a plurality. According to an exemplary embodiment, when a plurality of tuner units 1540 is formed, a plurality of broadcast signals may be output to a plurality of windows constituting a multi-window screen provided to the display unit 1515.

The tuner unit 1540 is implemented as an all-in-one with the display device 1500 or a separate device having a tuner unit electrically connected to the display device 1500 (for example, a set-top box (set-top box)). It may be implemented as a top box (not shown) or a tuner unit (not shown) connected to the input/output unit 1570).

The communication unit 1550 may connect the display device 1500 to an external device (eg, an audio device) under the control of the controller 1580. The controller 1580 may transmit/receive content to an external device connected through the communication unit 1550, download an application from the external device, or browse the web. Specifically, the communication unit 1550 may access a network and receive content from an external device (not shown).

As described above, the communication unit 1550 may include at least one of a short-range communication module (not shown), a wired communication module (not shown), and a mobile communication module (not shown).

In FIG. 15, a case where the communication unit 1550 includes one of a wireless LAN 1551, a Bluetooth communication unit 1552, and a wired Ethernet (Ethernet, 1553) is illustrated as an example.

Further, the communication unit 1550 may include a module combination including at least one of a wireless LAN 1551, a Bluetooth communication unit 1552, and a wired Ethernet (Ethernet, 1553). Further, the communication unit 1550 may receive a control signal from a control device (not shown) under the control of the controller 1580. The control signal may be implemented in a Bluetooth type, an RF signal type, or a WiFi type.

In addition to Bluetooth, the communication unit 1550 may further include other short-range communication (eg, near field communication (NFC), not shown), and a separate BLE module (bluetooth low energy, not shown).

The sensing unit (not shown) detects a user's voice, a user's image, or a user's interaction.

The microphone 1561 receives a user's uttered voice. The microphone 1561 may convert the received voice into an electric signal and output it to the controller 1580. The user voice may include, for example, a voice corresponding to a menu or function of the display apparatus 1500. For example, the recognition range of the microphone 1561 is recommended to be within 4 m from the microphone 1561 to the user's position, and the recognition range of the microphone 1561 is the size of the user's voice and the surrounding environment (eg, speaker sound, May vary in response to ambient noise).

The microphone 1561 may be implemented integrally with the display device 1500 or as a separate type. The separated microphone 1561 may be electrically connected to the display apparatus 1500 through a communication unit 1550 or an input/output unit 1570.

It will be readily understood by those of ordinary skill in the art that the microphone 1561 may be excluded depending on the performance and structure of the display device 1500.

The camera unit 1562 receives an image (eg, a continuous frame) corresponding to a user's motion including a gesture in the camera recognition range. For example, the recognition range of the camera unit 1562 may be within 0.1 to 5 m from the camera unit 1562 to the user. The user motion may include, for example, a user's face, an expression, a hand, a fist, a user's body part such as a finger, or a motion of a user part. The camera unit 1562 may convert the received image into an electric signal under the control of the controller 1580 and output it to the controller 1580.

The controller 1580 may select a menu displayed on the display apparatus 1500 using the received motion recognition result or may perform a control corresponding to the motion recognition result. For example, it may include channel adjustment, volume adjustment, and indicator movement.

The camera unit 1562 may include a lens (not shown) and an image sensor (not shown). The camera unit 1562 may support optical zoom or digital zoom by using a plurality of lenses and image processing. The recognition range of the camera unit 1562 may be variously set according to the angle of the camera and environmental conditions. When the camera unit 1562 includes a plurality of cameras, a 3D still image or a 3D motion may be received using the plurality of cameras.

The camera unit 1562 may be implemented integrally with the display device 1500 or as a separate type. A separate device (not shown) including the separated camera unit 1562 may be electrically connected to the display device 1500 through the communication unit 1550 or the input/output unit 1570.

It will be readily understood by those of ordinary skill in the art that the camera unit 1562 may be excluded according to the performance and structure of the display apparatus 1500.

The light receiving unit 1563 receives an optical signal (including a control signal) received from an external control device (not shown) through a light window (not shown) of the bezel of the display unit 1515. The light receiver 1563 may receive an optical signal corresponding to a user input (eg, touch, depress, touch gesture, voice, or motion) from a control device (not shown). A control signal may be extracted from the received optical signal under control of the controller 1580.

For example, the light receiving unit 1563 may receive a signal corresponding to a pointing position of a control device (not shown) and transmit it to the control unit 1580. For example, a user interface screen for receiving data or commands from the user through the display unit 1515 is output, and the user wants to input data or commands to the display device 1500 through a control device (not shown). In this case, the light receiver 1563 responds to the movement of the control device (not shown) when the user moves the control device (not shown) in a state in which a finger is in contact with a touch pad (not shown) provided on the control device (not shown). A signal to be transmitted may be received and transmitted to the controller 1580. Also, the light receiving unit 1563 may receive a signal indicating that a specific button provided in the control device (not shown) has been pressed and transmit the signal to the control unit 1580. For example, when the user presses a touch pad (not shown) provided in a button type on a control device (not shown) with a finger, the light receiving unit 1563 receives a signal indicating that the button type touch pad (not shown) has been pressed, and It can be transmitted to the control unit 1580. For example, a signal indicating that a button type touch pad (not shown) is pressed may be used as a signal for selecting one of the items.

The input/output unit 1570 controls video (eg, video), audio (eg, voice, music, etc.) and additional information (eg, video, etc.) from the outside of the display device 1500 under the control of the controller 1580. For example, EPG, etc.). The input/output unit 1570 is one of an HDMI port (High-Definition Multimedia Interface port, 1571), a component jack (1572), a PC port (PC port, 1573), and a USB port (USB port, 1574). It may include. The input/output unit 1570 may include a combination of an HDMI port 1571, a component jack 1572, a PC port 1573, and a USB port 1574.

It will be easily understood by those of ordinary skill in the art that the configuration and operation of the input/output unit 1570 may be variously implemented according to an embodiment of the present invention.

The controller 1580 controls the overall operation of the display apparatus 1500 and a signal flow between internal components (not shown) of the display apparatus 1500 and processes data. The controller 1580 may execute an OS (Operation System) stored in the memory 1590 and various applications when there is a user input or a preset and stored condition is satisfied.

The controller 1580 stores a signal or data input from the outside of the display device 1500, or a RAM (not shown) used as a storage area corresponding to various operations performed by the display device 1500, and a display device 1500. A control program for controlling) may include a ROM (not shown) and a processor (not shown).

The processor (not shown) may include a graphic processing unit (not shown) for processing graphics corresponding to video. The processor (not shown) may be implemented as a SoC (System On Chip) in which a core (not shown) and a GPU (not shown) are integrated. The processor (not shown) may include a single core, a dual core, a triple core, a quad core, and a multiple of the core.

In addition, the processor (not shown) may include a plurality of processors. For example, the processor (not shown) may be implemented as a main processor (not shown) and a sub processor (not shown) operating in a sleep mode.

The graphic processing unit (not shown) generates a screen including various objects such as icons, images, texts, etc. using an operation unit (not shown) and a rendering unit (not shown). The calculation unit calculates attribute values such as coordinate values, shape, size, color, etc. to be displayed for each object according to the layout of the screen by using the user interaction sensed through the detection unit (not shown). The rendering unit generates screens of various layouts including objects based on the attribute values calculated by the calculation unit. The screen generated by the rendering unit is displayed in the display area of the display unit 1515.

16 is a diagram illustrating a server and a display device for performing an AI model selection method based on input data according to the disclosed embodiment.

Referring to FIG. 16, the method of selecting an AI model based on input data according to the disclosed embodiment described with reference to FIGS. 1A to 11 may be performed through the display apparatus 1600 and the server 1650.

Here, the display device 1600 may correspond to the display device 1200, 1300, 1400, or 1500 according to the disclosed embodiment described with reference to FIGS. 12 to 15. In addition, the display device 1600 and the server 1650 may be connected through a communication network 1601, and the display device 1500 includes a communication unit (not shown) (not shown) (a communication unit shown in FIGS. 13 and 14). Through 1330, it is possible to access the communication network 1601.

The server 1650 may perform an operation according to a neural network. In addition, the display device 1600 and the server 1650 are, as a result of performing an operation according to the neural network, an AI model obtained by training a neural network, data required to train the AI model, and the AI model required to output the result. Data can be transmitted and received.

The AI model mounted in the display apparatus 1600 illustrated in FIG. 16 may be the on-device AI model described in FIG. 9, and the AI model mounted in the server 1650 may be a server-based AI model.

Operations performed by the display apparatus 1600 and the server 1650 will be described in detail below with reference to FIGS. 17 to 19. In Figs. 17 to 19, the same configuration as in Fig. 16 is illustrated using the same reference symbols.

17 is a block diagram specifically showing the server illustrated in FIG. 16.

Referring to FIG. 17, a server 1650 according to an embodiment may include a DB 1710, a communication unit 1720, and a processor 1730. The server 1650 may operate in conjunction with the display device according to the disclosed embodiment, and may perform any one of input data analysis, object recognition, object classification, and result data output by performing an operation according to a neural network. have.

The DB (database) 1710 may include a memory (not shown), and stores at least one of at least one instruction, program, and data necessary for the server 1650 to perform a predetermined operation in the memory (not shown). I can.

In addition, the database (DB) 1710 may store data necessary for the server 1650 to perform an operation according to a neural network. Specifically, the DB (database) 1710 may store a training data set (eg, 300 in FIG. 3) necessary to train the AI model. In addition, the DB 1710 may store a plurality of AI models. Also, the DB 1710 may store the misclassification data acquired in step S210. In addition, the DB 1710 may store additional AI models, train additional AI models on misclassified data, and store the trained additional AI models.

The communication unit 1720 may include one or more components that enable communication with the display device 1600. Since the specific configuration of the communication unit 1720 corresponds to the configuration of the communication unit 1330 described in FIG. 13, detailed descriptions are omitted.

The processor 1730 controls the overall operation of the server 1650. For example, the processor 1730 may generally control the DB 1710 and the communication unit 1720 by executing programs stored in the DB 1710 of the server 1650. The processor 1730 may perform some operations of the image display device 100 in FIGS. 1A to 12 by executing programs stored in the DB 1710.

The processor 1730 performs an operation required for training an AI model, outputting a result of the AI model, and transmitting the result output from the AI model to the display device 1600, or an operation required for transmitting and receiving data with the display device 1600. Can be done.

Hereinafter, operations performed in the server 1650 and the display device 1600 will be described in detail with reference to FIGS. 18 and 19. In addition, in Figs. 18 and 19, the same configuration as in Figs. 16 to 17 is shown using the same reference symbols.

18 is a block diagram illustrating a display device communicating with a server.

FIG. 19 is a diagram for describing operations performed in the server and display device illustrated in FIG. 16. Specifically, in FIG. 19, operations performed in the server 1650 and operations performed in the display apparatus 1600 are illustrated using a flowchart.

In FIG. 18, for convenience of explanation, the processor 1210 included in the display apparatus 1600 is shown as a'first processor', and the processor 1730 included in the server 1650 is shown as a'second processor'. .

18 and 19, an AI model selection method 200, 1000, or 1100 based on input data according to one or another embodiment described with reference to FIGS. 1A to 11 is a display apparatus 1600 and a server ( 1650).

Specifically, the DB 1710 may store a training data set (eg, the training data set 300 shown in FIG. 3) and a plurality of AI models.

18 and 19, the second processor 1730 executes at least one instruction and uses the training data set (eg, 300) stored in the DB 1710 to each of a plurality of AI models. Train (S1911). Here, since the operation of step S1911 is the same as that of step S1005 described in FIG. 10, detailed descriptions are omitted. Then, as a result of performing step S1911, the second processor 1730 may acquire misclassification data (S1912). Here, since the operation of step S1912 corresponds to the step S210 described in FIG. 2A, a detailed description will be omitted. In addition, the obtained misclassification data may be stored in the DB 1710 under the control of the second processor 1730.

In addition, the DB 1710 may store an additional AI model. Here, the DB 1710 may store both the additional AI model before the training is performed and the additional AI model after the training is performed.

Subsequently, the second processor 1730 executes at least one instruction and trains each of the plurality of AI models by using the misclassification data stored in the DB 1710 (S1913). Here, since the operation of step S1913 corresponds to the step S220 described in FIG. 2A, a detailed description is omitted.

In general, multiple AI models can exist in a wide variety of forms, and a training data set must be formed as a set of many data for training. Therefore, it is preferable that the plurality of AI models and training data sets are stored and processed in the server 1650 having a large storage capacity. Accordingly, the operation of training a plurality of AI models may be performed in the server 1650 connected to the display apparatus 1600 and the communication network 1601.

The display apparatus 1600 may receive an additional AI model for which training has been completed from the server 1650. Specifically, under the control of the first processor 1210, the communication unit 1330 may receive an additional AI model that has been trained in step S1913 through the communication unit 1720 of the server 1650. In addition, the first processor 12100 may store the received additional AI model in the memory 1440.

The first processor 1210 inputs first data as an additional AI model (S1951), and outputs misclassification probabilities for each of the plurality of AI models (S1952). Here, since the operation of step S1952 corresponds to the step S230 described in FIG. 2A, a detailed description will be omitted.

Then, the first processor 1210 selects any one of the plurality of AI models mentioned in step S1911 based on the misclassification probabilities obtained in step S1952 (S1953). Here, since the operation of step S1953 corresponds to the step S240 described in FIG. 2A, a detailed description will be omitted.

Subsequently, the first processor 1210 may output a result corresponding to the first data by using the AI model selected in step S1953 (S1954). Here, since the operation of step S1954 corresponds to the step S1050 described in FIG. 10, detailed descriptions are omitted.

That is, in FIG. 19, a plurality of AI models trained by the training data set in S1911 may be a server-based AI model. In addition, the additional AI model in the S1913 stage can also be a server-based AI model. The additional AI model trained in step S1913 and transmitted to the display apparatus 1600 may be operated as an on-device AI model after being received and stored by the display apparatus 1600.

In general, the display apparatus 1600 may have limited memory storage capacity, operation processing speed, and ability to collect a learning data set compared to the server 1650. Accordingly, an operation requiring a large amount of data and a large amount of computation may be performed by the server 1650, and then necessary data and an AI model may be transmitted to the display device 1600 through a communication network. Then, the display apparatus 1600 can perform the required operation quickly and easily by receiving and using the necessary data and the AI model through the server without a processor having a large memory and a fast computation capability.

In addition, FIG. 19 shows an example in which the server 1650 performs the training operation (S1911) of a plurality of AI models, the misclassification data acquisition operation (S1912), and the training operation of the additional AI model (S1913). In addition to this, the server 1650 may directly perform at least one of an operation of outputting misclassification probabilities (S1952), an operation of selecting an AI model (S1953), and an operation of outputting a result (S1954). In addition, data corresponding to the result to be performed may be transmitted to the display apparatus 1600.

In addition, the display apparatus 1600 has been shown as an example of performing an operation of outputting misclassification probabilities (S1952), an operation of selecting an AI model (S1953), and an operation of outputting a result (S1954), but the display apparatus 1600 ) May directly perform at least one of a training operation of a plurality of AI models (S1911), a misclassification data acquisition operation (S1912), and a training operation of an additional AI model (S1913).

Depending on the design specifications, data processing capacity, memory capacity, data processing speed, etc. of the display apparatus 1600 and the server 1650, operations included in the method of selecting an AI model based on input data according to the disclosed embodiment may be performed. ) And the server 1650 may be determined whether or not to be performed. For example, operations that use large amounts of data and require fast data processing may be performed by the server 1650, and relatively simple operation operations or operations requiring security may be performed by the display device 1600. .

The AI model selection method based on input data according to one or another embodiment of the present disclosure may be implemented in the form of program instructions that can be executed through various computer means and recorded in a computer-readable medium. In addition, an embodiment of the present disclosure may be formed as a computer-readable recording medium in which one or more programs including instructions for executing an AI model selection method based on input data are recorded.

The computer-readable medium may include program instructions, data files, data structures, and the like alone or in combination. The program instructions recorded in the medium may be specially designed and configured for the present invention, or may be known and usable to those skilled in computer software. Examples of computer-readable recording media include magnetic media such as hard disks, floppy disks, and magnetic tapes, optical media such as CD-ROMs and DVDs, and magnetic media such as floptical disks. -A hardware device specially configured to store and execute program instructions such as magneto-optical media, and ROM, RAM, flash memory, and the like. Examples of the program instructions include not only machine language codes such as those produced by a compiler, but also high-level language codes that can be executed by a computer using an interpreter or the like.

In addition, the method for selecting an AI model based on input data according to the disclosed embodiment includes: obtaining a sentence composed of multiple languages; And using a multilingual translation model, obtaining vector values corresponding to each of the words included in the sentence composed of the multilingual, converting the obtained vector values into vector values corresponding to the target language, and the converted It may be implemented as a computer program product including a recording medium in which a program for performing an operation of obtaining a sentence composed of the target language based on vector values is stored.

Although the embodiments have been described in detail above, the scope of the present invention is not limited thereto, and various modifications and improvements by those skilled in the art using the basic concept of the present invention defined in the following claims are also the scope of the present invention. Belongs to.

1200, 1300, 1400, 1500: display device
1210: processor
1220: display
1330: Ministry of Communications
1440: memory
1450: user interface

Claims (20)

When each of a plurality of AI (Artificial Intelligence) models receives a training data set, classifies it into at least one class, and outputs the result, the data corresponding to the misclassified result of the training data set Obtaining misclassified data;
Learning the misclassified data from an additional AI model to obtain a learned additional AI model;
Outputting misclassification probabilities for each of the plurality of AI models when first data is input to the learned additional AI model; And
And selecting any one of the plurality of AI models based on the misclassification probabilities. The method of claim 1, wherein the selecting step
And selecting a model corresponding to a misclassification probability having the smallest value among the misclassification probabilities among the plurality of AI models. The method of claim 1, wherein the learning step
Learning from which of the plurality of AI models the misclassified data is classified data from an additional AI model receiving the misclassified data; And
Comprising the step of obtaining the learned additional AI model, AI model selection method based on input data. The method of claim 3, wherein the learning step
AI model selection based on input data, including the step of learning, by the additional AI model receiving the misclassification data, characteristics of data that cannot be learned by each of the plurality of AI models and distribution of data Way. The method of claim 1,
The method of selecting an AI model based on input data, further comprising the step of receiving first data from the selected AI model and outputting a result of classifying it into the at least one category. The method of claim 1, wherein the additional AI model
Data processing method using an AI model, which is an AI model different from the plurality of AI models. The method of claim 1, wherein the misclassification probability is
An AI model selection method based on input data indicating a probability that each of the plurality of AI models incorrectly classifies the input data. The method of claim 1, wherein obtaining the misclassification data
Receiving the training data set from each of the plurality of AI models, classifying the training data set into the at least one class, and outputting the result; And
A method for selecting an AI model based on input data, comprising the step of acquiring the misclassified data, which is data having an incorrect value among the training data sets. The method of claim 1, wherein the plurality of AI models
A method of selecting an AI model based on input data in which at least one of a hyper-parameter, a model structure, a training technique, and a data set input for training is formed differently. The method of claim 1,
The input further comprising adjusting a plurality of weight values applied to each of the plurality of AI models in an ensemble AI model formed by combining the plurality of AI models based on the misclassification probabilities How to choose an AI model based on data. The method of claim 10, wherein adjusting the plurality of weight values
A method for selecting an AI model based on input data, comprising adjusting a plurality of weight values applied to each of the plurality of AI models so that an applied weight value decreases when the misclassification probability is high. The method of claim 10,
The method of selecting an AI model based on the input data, further comprising inputting the first data into the ensemble AI model and outputting a final classification result corresponding to the first data. The method of claim 1, wherein the plurality of AI models
AI model selection method based on input data, including an ensemble AI model formed by combining a plurality of AIs. The method of claim 1, wherein the plurality of AI models
AI model selection method based on input data, including at least one on-device AI model and at least one server based AI model. display; And
By executing at least one instruction,
When each of a plurality of AI (Artificial Intelligence) models receives a training data set, classifies it into at least one class, and outputs the result, the data corresponding to the misclassified result of the training data set Misclassified data is acquired, and an additional AI model is trained as the misclassified data to obtain a learned additional AI model, and when first data is input as the learned additional AI model, each of the plurality of AI models A display apparatus comprising a processor that controls to output misclassification probabilities for the output, and selects any one of the plurality of AI models based on the misclassification probabilities. The method of claim 15, wherein the processor
A display device for selecting a model corresponding to a misclassification probability having a smallest value among the misclassification probabilities among the plurality of AI models. The method of claim 15, wherein the processor
In the additional AI model receiving the misclassification data, controlling to learn from which model the misclassified data is data classified from among the plurality of AI models. The method of claim 15, wherein the processor
The display device, wherein the training data set is input to each of the plurality of AI models, and the result obtained by classifying the training data set into the at least one class is output. The method of claim 15,
Further comprising a communication unit in communication with the server,
The processor is
Controlling the misclassification data to be received from the server through the communication unit. In a computer-readable recording medium,
When each of a plurality of AI (Artificial Intelligence) models receives a training data set, classifies it into at least one class, and outputs a result, data corresponding to a misclassified result of the training data set Obtaining misclassified data;
Learning the misclassified data from an additional AI model to obtain a learned additional AI model;
Inputting first data into the learned additional AI model and outputting misclassification probabilities for each of the plurality of AI models; And
A recording medium storing a program for performing an operation of selecting any one of the plurality of AI models based on the misclassification probabilities.

Images